Finding Boundaries in Images

Abstract

In computational vision, finding the boundaries of the regions in a image which correspond to different surfaces in the scene is usually approached as a problem of detecting brightness edges. In this paper, we argue that this is a limited view. Boundaries in images could be associated with differences on a number of visual attributes-brightness, color, texture, stereoscopic disparity and motion-all of which are utilized in human vision. Machine vision systems should do the same. We argue that convolution of the image with a bank of Gaussian derivative filters is a suitable common first stage for this task. We also present some new results on the problem of detecting and localizing brightness edges composed of step, peak and roof profiles.